Signal-to-Noise Ratio for Voice Activity Detection Devices

How to Design in VAD

Users needs to be able to give AI glasses voice commands and be sure the device is only picking up their speech, not other sounds. Vibration from the user’s voice is the key to the next evolution of voice commands, one that engineers can tap into with accelerometer technology.

Better Hearing with Good Vibrations

The dilemma of how to make voice commands functional and safe in noisy weather such as wind, loud crowds, and busy streets, found its answer in nature. Ossicles, the three smallest bones in the human body, work together to transmit and amplify sound vibrations from the eardrum to the inner ear. But other bones also vibrate with speech. Human skulls and jawbones vibrate when someone speaks. So, unlike a microphone that will pick up surrounding sound from non-user voices and environmental noise like traffic and wind, AI glasses use vibration along the skull and jawbones to trigger listening only when the user is talking. Known as Accel-Based Voice Activity Detection (VAD), this ability to determine when the user is speaking using the vibration induced by the user’s Voice comes with many benefits.

Advantages of Accelerometer Voice Activity Detection

Effective Accel-Based Voice Activity Detection, also called Voice Vibration Detection (VVD), can improve user experience significantly. For example, earbuds with VAD can sense the vibrations of a user speaking and can pause music or switch from noise cancelling to transparency/hear-through mode. This allows users to seamlessly jump from listening to conversation.

Such accelerometer-based VAD can also increase Keyword Spotting (KWS) performance by increasing microphone performance. The VAD helps a device’s wake-up host activate to run KWS or Automatic Speech Recognition (ASR). Combining accelerometer VAD and microphone VAD lowers the False Accept Rate (FAR) by making sure it’s tracking a user’s requests. Waking with vibration also helps ensure that the user’s voice is heard. This feature improves voice ID security, keeping commands secure while avoiding spoofs.

When a user speaks, the accelerometer in a device’s Inertial Measurement Unit (IMU) detects the vibrations in the head generated by the voice. Understanding the different steps the signal path takes from voice vibrations on the head to the signal that activates algorithms on the device as shown in Figure 1 will help improve the quality of the signal and the final performance the user experiences.

Signal Position and Device Design

Speech vibration levels vary based on head position, and the resulting skull vibrations are transmitted to the device through an interface. The quality of this interface heavily influences the level of vibrations transmitted from the head to the device.

For example, the silicone part of earbuds acts as a soft spring, creating a low-pass filter effect that typically removes vibrations above 500Hz. The device’s mass will also influence the transmission, so a heavier earbud will further lower the cutoff frequency of the low-pass filter. This filtering can be modeled as a transfer function, which describes how much vibration is passing through the interface from the skin to the accelerometer at each frequency.

Positioning is important to maximize the amplitude of the voice signal, as well as to keep the transfer function high; devices must ensure good skin contact, applying light pressure and avoiding soft material between skin and accelerometer. To ensure good vibration transmission for smart glasses, the accelerometer should be placed close to the temple bars with a force above 1N between temple bars and head.

User Motion and Sensor Noise

Two other features can help device design to improve VAD function. When the user moves or touches the device, it creates vibrations that are captured by the IMU. Good device design can help minimize unwanted vibrations, along with false triggers that can reduce the performance of VAD algorithms. Those vibrations may also hide the voice vibration, causing missed voice events.

Sensors add a white noise, constant at all frequencies, to the acceleration at the IMU location, which reduces the performance by hiding small amplitude details. The higher the sensor noise the lower the potential performance of the Accel-based VAD.

Optimizing Signal-to-Noise Ratio in VAD Design

Signal-to-Noise Ratio (SNR) defines the relative amplitude of speech compared to the total background noise—the sum of sensor noise and vibrations caused by user motion.

The measurement of speech level over sensor noise shown in Figure 2 showcases the signal to noise ratio (SNR) for TDK’s IMU with varying force between the head and the smart glasses frames for three different users.

To check the integration of the IMU, engineers must estimate the SNR when the user is not moving. To estimate the SNR, follow these steps:

Compute the root mean square (RMS) of the IMU background noise (high-pass filtered at 80Hz). A minimum duration of 1s is recommended to compute the RMS.

Compute the RMS of the IMU signal during speech (high-pass filtered at 80Hz).

Compute the ratio as shown in Figure 3.

A higher SNR means speech signals are much louder than the background noise, while an SNR close to 0dB means the speech is nearly imperceivable in the measured signal. The SNR depends on the whole signal path from head vibrations to the IMU.

By counting the number of speech events for each SNR value and comparing them to correctly classified events and missed speech events, it’s clear that the VVD algorithm has good performance if the speech SNR is above 5dB as shown in Figure 3. If the user speaks quietly, if the sensor is placed poorly on the device, if poor coupling between head and device exists, or if the device is poorly designed, the SNR may be below 5dB where the algorithm fails to detect most words.

Optimized SNR Delivers Better VAD

With a better understanding of how SNR affects VAD quality, as well as guidance on how to achieve proper SNR, engineers can fine-tune smart glasses designs for maximum performance of waking on command and other features. By tapping into the benefits of accelerometer VADs with an algorithm like TDK’s VVD, smart glasses will be able to achieve more than any microphone could on its own to deliver a superior user experience.

RESOURCES
TDK | www.tdk.com

PUBLISHED IN CIRCUIT CELLAR MAGAZINE • JULY 2026 #432 – Get a PDF of the issue

Sponsor this Article